An attention mechanism-based hybrid TimeAttentionBiLSTM architecture for long-term traffic forecasting

Abstract

Long-term traffic flow forecasting in real time has become an important research problem due to the rapid development of public transport facilities. High-density transportation features of traffic flow forecasting provide a convenient, fast, accurate, and comfortable boarding environment for the public. The change in the traffic flow is an important indicator for public transportation to provide facilities for passengers. In this paper, we propose an attention mechanism-based hybrid TimeAttentionBiLSTM architecture for long-term traffic forecasting. The TimeAttentionBiLSTM is able to capture periodic and nonperiodic patterns of the temporal features of traffic data. This architecture uses the bidirectional long short-term memory as an encoder and gated recurrent unit as a decoder to capture the long-term traffic dependencies of the traffic patterns. The TimeAttentionBiLSTM focuses on the important features of time series due to the attention mechanism. The TimeAttentionBiLSTM architecture improves the performance of traffic forecasting for the granularities of 12-h, 24-h, 48-h, and 72-h prediction. The experiments conducted on the Madrid city traffic data show the superiority of the TimeAttentionBiLSTM over nine other state-of-the-art baseline approaches in terms of mean average error, accuracy, and root-mean-square error evaluation metrics. © 2025 Elsevier B.V., All rights reserved.